Tuner

Thomas Torsney-Weir

VDA research group, University of Vienna

Acknowledgements

Torsten Moeller
Ahmed Saad
Britta Weber
Steven Bergner
Hans-Christian Hege
Jean-Marc Verbavatz

Image segmentation

Algorithms

  • Thresholding
  • Snakes
  • Watershed
  • Graph cuts
  • Variational methods

Parameters

  • Noise filtering
  • Regularization
  • Edge thresholds
  • Termination conditions

Parameters

Picking parameters

Picking parameters

Objective measures

Problem characterization

Users
Segmentation algorithm developers
Data
Segmentation algorithm (continuous)
  • inputs: parameters (several)
  • outputs: objective measures (several)
Tasks
  • "Best" parameter setting
  • Range of possible performance
  • Tradeoffs amongst objective measures
  • Sensitivity of parameters
  • How many simulations to run?

Manual pipeline

Manual pipeline

Tuner

Before Tuner After Tuner
Workflow Manual
Process Time consuming
Confidence Low

Tuner

Before Tuner After Tuner
Workflow Manual Systematic
Process Time consuming Fast
Confidence Low High

Visual parameter space exploration

conceptual pipeline

Michael Sedlmair, Christoph Heinzl, Stefan Bruckner, Harald Piringer, and Torsten Möller "Visual parameter space analysis: A conceptual framework" IEEE Transactions on Visualization and Computer Graphics. 20(12) 2014.

Visual parameter space exploration

image segmentation pipeline

Sampling

tuner pipeline

Reconstruction

tuner pipeline

Exploration

tuner pipeline
Tuner interface
Response view - tuner interface
Response view
Pareto view - tuner interface
Pareto view
View controls - Tuner interface
View controls
Histograms - Tuner interface
Histograms
Controls - Tuner interface
Controls
Response view blowup
Colormap - open
Colormap - filtered
Response view blowup
Pareto view blowup
Pareto view points band
Pareto view p1
Pareto view p2

Uncertainty views

Prediction

Error view
Error view

Optimization

Gain view
Gain view

Predictor error

Prediction error from jones paper

Donald R. Jones, Matthias Schonlau, and William J. Welch "Efficient global optimization of expensive black-box functions." Journal of Global Optimization. 13(4) 1998.

Predictor error

Expected gain

Exp gain img

Donald R. Jones, Matthias Schonlau, and William J. Welch "Efficient global optimization of expensive black-box functions." Journal of Global Optimization. 13(4) 1998.

Expected gain

before
after

Donald R. Jones, Matthias Schonlau, and William J. Welch "Efficient global optimization of expensive black-box functions." Journal of Global Optimization. 13(4) 1998.

Expected gain

Systematic workflow

Before Tuner After Tuner
Workflow Manual Systematic
Process Time consuming Fast
Confidence Low High

Integration

sampling interface
sampling interface
  • Tuner runs user-defined executable
  • CSV interface

Integration

sampling interface
  • Inputs have upper/lower range
  • Outputs are automatically determined

Integration

Integration

Integration

Tuner pipeline

Pipeline img

Resampling

gain view
gain view
gain view
gain view
gain view
gain view

High confidence

Before Tuner After Tuner
Workflow Manual Systematic
Process Time consuming Fast
Confidence Low High

Sensitivity

What parameter settings are sensitive?

sens img
sens img

Different datasets

What parameters matter at different noise levels?

No noise
No noise
10λ noise
10λ noise

Thanks!

Tuner img

thomas.torsney-weir@univie.ac.at

http://tuner.cs.univie.ac.at/